Key Idea: Machine learning finds patterns in data instead of following rules a programmer wrote. What a model can learn is decided entirely by the data it is given — which is why so many questions on this topic are really about the data.
Paper 1
- Define the three types; identify which suits a scenario.
Paper 2
- Extended ethical reasoning about a deployed system.
Both
- Ethics marks come from consequences, not from naming a concern.
🧠 The three types
| Type | Learns from | Use it when |
|---|---|---|
| Supervised | Labelled examples | The correct answers already exist — spam, diagnosis |
| Unsupervised | Unlabelled data | The groupings are what you are trying to discover |
| Reinforcement | Rewards for outcomes | Only the outcome can be scored — a game, a controller |
Ask: do the correct answers already exist? If yes, supervised. If the groupings are the thing being discovered, unsupervised. If only the end result can be scored, reinforcement.
📊 Why accuracy misleads
Accuracy is one number, and it hides which predictions are wrong — which is usually what matters.
Important: On a condition affecting one person in a thousand, a model that answers 'no' every time is 99.9% accurate and has never identified a single case. The errors that matter are invisible in the single figure.
What to report instead
- False positives and false negatives separately — they carry very different costs
- Accuracy per group, not just overall, or a failure on a small group is hidden
- For medical screening a false negative is far worse: the condition goes untreated
🖥️ Training against inference
| Training | Inference | |
|---|---|---|
| Work | The whole dataset, repeatedly | One input, one pass |
| Adjusts parameters? | Yes | No — they are fixed |
| Takes | Hours to weeks | Milliseconds |
| Needs | Many accelerators | Sometimes just a phone |
Training is dominated by matrix arithmetic, and every element can be computed independently — so thousands of simple cores beat a few powerful ones. But sequential, branch-heavy code gains nothing from parallel hardware.
📝 Exam-style questions
A supermarket wants to group its customers by shopping habits, without deciding the groups in advance. Identify the type of machine learning required and justify your choice.
🔒 Model answer plan
See the mark-by-mark plan — for / against / judgement, with marking guidance — in study mode.
A screening model is tested on 4000 scans. It correctly identifies 76 of the 100 that show the condition, and correctly clears 3850 of the 3900 that do not. Determine its accuracy, and explain why that figure is misleading.
🔒 Model answer plan
See the mark-by-mark plan — for / against / judgement, with marking guidance — in study mode.
✅ Quick check
Cover the answers.
Which type learns from rewards rather than answers? Reinforcement learning.
What is overfitting? Fitting the training data so closely — including its noise — that new data is handled badly. The model learned the examples, not the pattern.
Why keep test data separate from training data? A model can score well on data it memorised. Held-back data measures generalisation, which is what actually matters.
Why does a GPU suit training? Training is parallel matrix arithmetic, and a GPU has thousands of cores to do it at once.
Can a phone run a model it could not train? Yes — inference is one forward pass with fixed parameters, a tiny fraction of training's cost.
Exam tips
- Ask whether labels exist — that decides supervised versus unsupervised immediately.
- Never quote accuracy alone; say what a majority-class guesser would score.
- False negatives and false positives carry different costs. Name which matters here.
- Training adjusts parameters over the whole dataset; inference is one pass with them fixed.
- Parallel hardware only helps parallel work — sequential code gains nothing from a GPU.
- Overfitting is fixed by more varied data or a simpler model, not by training longer.